Debugging the Work Behind the Diff: Observability for AI Coding Agents

The talk is now being edited and going to be available soon to multipass and full ticket holders
Registered portal users get access 15 days after and the rest 30 days after
Bookmark
Rate this content

After a weekend away at a cabin with limited cell service, I returned to work and discovered unexpected Fable usage (50% of my weekly budget was blown before I even got started). Tracing the consumption back into the contributing session revealed Fable checking on the status of a PR every 5 minutes, death by a thousand cuts. Icing on the cake is the pull request contained only one Markdown file: a plan for a feature I wasn't sure I wanted. I found the problem because I happened to look. What could have exposed it sooner?

This talk follows that investigation from an unexpected usage pattern to an actionable alert. We will examine the different properties of the same incident: accumulated consumption, repeated scheduling, unattended wakeups, premium-model use, and the cost of each check-in. Each suggests a different detection strategy, with different strengths and limitations. A workflow can consume resources unnecessarily even when its individual commands succeed, while a long session or a lack of file edits does not automatically mean the work has no value.

This talk has been presented at AI Coding Summit NYC, check out the latest edition of this Tech Conference.

Elliot Gardiner
Elliot Gardiner
17 Nov, 2026
Video transcription, chapters and summary will be available later.